In the cancer centre of a large public hospital in New Zealand, patients came in for chemotherapy and then sat and waited for their drugs. A survey of 62 patients found waits of between one and four hours. The average was 129 minutes.
The staff weren’t idle. In the week before anyone tried to fix it, nurses had worked 61 hours of overtime, and the pharmacy that made the drugs had a backlog of up to 340 minutes. Its workload swung wildly, the researchers who studied it wrote, “with idle staff looking for work one minute, only to have to work overtime later.”
The obvious answers were more pharmacists or a faster production unit. The hospital did neither.
The slowest hiker
In 1984 a physicist named Eliyahu Goldratt published a business book in the form of a novel. In The Goal, a struggling plant manager, Alex Rogo, takes his son’s Boy Scout troop on a ten-mile hike. At two miles an hour it should take five hours.
The boys start together and quickly spread out along the trail. At a trail marker, Alex works out that they have been covering only one mile an hour. The fast boys race ahead. At the back, a boy named Herbie keeps falling behind, and everyone behind him can go no faster than he does.
So Alex stops the troop and turns the line around. Herbie goes to the front, the fastest boys to the back. Now the troop stays together. Then they take some of the load out of Herbie’s pack and share it out among the others. With less to carry, Herbie speeds up, and so does everyone. From then on, the troop holds its two miles an hour.
That’s the theory of constraints in a hiking boot. A line can turn out only as much as its tightest step can handle. Speed up any other step and the work just piles up in front of that one, like the gap opening between the fastest hikers and Herbie.
Practitioners of the theory teach the lesson as five steps. Find the constraint. Get everything you can out of it. Make everything else work to its pace. Only then pay to expand it. And when it stops being the constraint, start again, because something else now is. The TOC Institute warns that people tend to skip straight to the fourth step, the one that costs money.
The constraint was a rule
The New Zealand hospital found its Herbie in the pharmacy. But when the team looked closer, the slow part wasn’t the people mixing the drugs. It was a rule about when they were allowed to start.
A patient’s drugs were made from scratch only after a specialist confirmed that day, from blood tests and usually a clinic visit, that the patient was fit for treatment. The belief was that this saved money. Chemotherapy drugs are hugely expensive, and a dose made for a patient who can’t be treated that day might be thrown away.
Goldratt himself had come to see that policies can be constraints, and harder to shift than machines. The hospital tested that. For four weeks, the pharmacy made the drugs in advance, most of them the day before.
The longest wait fell from four hours to 30 minutes. Most patients didn’t wait at all. Nursing overtime dropped by two-thirds. And the pharmacy, which had been the bottleneck, made 37% more items for the cancer centre during the trial than its pre-trial workload projected, with no extra overtime.
The trial had a cost. The cost of wasted drugs rose from 0.4% to 4%. So the team kept the specialist check for the 10% of drugs that were most expensive and made the other 90% ahead of time. Waste fell back close to 0.4%. No patient survey was done after that, but the researchers later estimated average waits at about 16.5 minutes, down from 129. A nurse manager who joined later said he was “initially really surprised to see all the chemo waiting.”
A faster hiker at the front
Imagine a team lead who gives five engineers an AI coding assistant. By the end of the month they are writing far more code. The same two senior engineers still review all of it, and the queue of changes waiting for review is longer every Friday.
That is roughly what the analytics firm Faros AI reported in 2025, from the work records of more than 10,000 developers on 1,255 teams. On teams that used AI heavily, developers merged 98% more pull requests, the batches of code changes that need a reviewer’s approval. Review time on those pull requests rose 91%.
The data shows patterns, not causes, and it comes from a company that sells measurement tools. But its own diagnosis reads like Goldratt. Across company-wide throughput and quality, it found no significant link to AI adoption. The bottleneck, it said, was “human approval.”
The coding assistant had made the fastest hikers faster. Herbie was still at the back, carrying a heavier pack than before.